void CompareNumericalAndAutodiffGradients(
    Model model, Model grad_model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, bool use_function,
    double abs_error = 1e-2);

void CheckTensorValue(AbstractTensorHandle* t, absl::Span<const float> manuals,
                      absl::Span<const int64_t> dims, double abs_error = 1e-2);

Model BuildGradModel(Model forward, GradientRegistry registry);

//   outputs = tf.function(model)(inputs)
// else:
//   outputs = model(inputs)
Status RunModel(Model model, AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs, bool use_function);

Status BuildImmediateExecutionContext(bool use_tfrt, AbstractContext** ctx);

// Return a tensor handle with given type, values and dimensions.

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will

//    the SavedModel.
// - `graph` must be a graph newly allocated with TF_NewGraph().
//
// If successful, populates `graph` with the contents of the Graph and
// `meta_graph_def` with the MetaGraphDef of the loaded model.
TF_CAPI_EXPORT extern TF_Session* TF_LoadSessionFromSavedModel(
    const TF_SessionOptions* session_options, const TF_Buffer* run_options,
    const char* export_dir, const char* const* tags, int tags_len,

// set, proper uses configuration set by user.
typedef struct TP_OptimizerConfigs {
  size_t struct_size;
  void* ext;  // reserved for future use
  TF_TriState disable_model_pruning;
  TF_TriState implementation_selector;
  TF_TriState function_optimization;
  TF_TriState common_subgraph_elimination;
  TF_TriState arithmetic_optimization;
  TF_TriState debug_stripper;

  TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// Sets the default execution mode (sync/async). Note that this can be
// overridden per thread using TFE_ContextSetExecutorForThread.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetAsync(TFE_ContextOptions*,

#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Abstract interface to an operation.
// This interface allows building and executing an operation in either
// tracing or immediate execution mode.
class AbstractOperation {
 protected:
  enum AbstractOperationKind {
    kGraph,
    kMlir,
    kEager,
    kTfrt,
    kTape,
    kOpHandler
  };

// Core APIs
// -----------------------------------------------------------------------------

// A TF_ExecutionContext stores knowledge about how to execute an operation.
// E.g. it could know whether we're in eager mode or graph mode, keeps track
// of gradient tapes, etc.
typedef struct TF_ExecutionContext TF_ExecutionContext;

// A TF_AbstractTensor is an input to an operation. E.g. it could be a union